Research On Radar Emitter Deinterleaving And Classification With Machine Learning

Radar emitter deinterleaving and recognition have always been a hot research issues in the field of electronic reconnaissance.However,the regularity of the radar pulse sequence has been seriously damaged because of three reasons: 1.the development of radar low interception probability technology;2.the complexity of the working system;3.the comprehensive application of anti-jamming technology.These phenomena make the signal-to-ratio very low and the ratios of aliasing pulses are high,which make the traditional deinterleaving and recognition methods driven by the model face serious challenges.Machine learning is a data-driven signal processing and information acquisition tool.In recent years,it has been widely used in image processing,data mining,natural language processing,speech recognition and other fields and has achieved great success.Because the data-driven method has good adaptability to large sample data containing errors,and can meet the needs of real-time processing after early training,it is introduced to the field of radar reconnaissance processing to solve traditional radar electronic reconnaissance processing fields to deal with the problem of weak adaptability to complex environments.This paper focuses on the application of machine learning technology in radar pulse deinterleaving and recognition in complex electromagnetic environments.The main work and innovative results are summarized as follows:(1)For the problem of rapid deinterleaving of important targets with known parameters,considering the high proportion of missing pulses and spurious pulses and large measurement errors in a complex electromagnetic environment,starting from the radar aliasing model of pulsed radiation sources,the autoencoder is used to compress and reconstruct the information,an unsupervised pulse sequence denoising method based on autoencoder and a deinterleaving method based on denoising autoencoder are proposed.In the process of denoising,the use of encoder is to extract useful information,and the decoding process to reconstruct the original information to achieve the suppression of false pulses,repair the missing pulses and correction of measurement errors.This method does not require the original pulse sequence as label data,but directly analyzes and extracts the common features between noisy data as unsupervised processing.Compared with traditional denoising methods,it has a stronger ability to adapt to noise.The fast deinterleaving method based on the noise reduction autoencoder actively add noise as the network input,and extracts the robust effective features layer by layer from the Time-of-Arrival(TOA)sequences through the autoencoder for deinterleaving.The simulation experiments prove that compared with the traditional method,the new method has obvious advantages such as simple model,fast convergence speed and strong robustness.(2)Aiming at the conventional pulse deinterleaving problem,a new deinterleaving structure based on Iterative Convolutional Neural Networks(ICNN)is proposed.The iterative unit in the network structure deletes the single pulse sequence output from the sorting from the original overlapping pulse sequence each time,so as to simplify the original sequence and realize iterative deinterleaving.ICNN not only can output the results one by one,but also can handle the sorting problem with uncertain number of target sources.In addition,because the order of the output target and the sample target cannot be determined during the sorting process,this paper proposes a multi-target training(MTT)method,which enables ICNN to solve the problem of matching the target and output during the training process.Simulation experiments show that under the same parameter settings,ICNN has higher deinterleaving accuracy and stronger adaptability to errors than traditional methods.(3)Facing the problem of high ratios of lost and spurious pulses,this paper proposes a CNN-PRIR(Convolutional Neural Networks based Pulse Repetitive Interval Recognition)based on convolutional neural network to solve the problem of PRI recognition.This method uses the powerful representation ability of deep learning to establish the mapping relationship between the radiation source pulse and the repetition frequency type.There are many types of PRI of the radiation source,and it is difficult to be accurately recognized in a low signal-to-noise ratio environment.However,due to the periodicity and locality of the repetition sequence,the deep convolutional neural network can extract the local features of the input,so it is very suitable for solving the problem of PRI type identification.Simulation experiments show that the CNN-PRIR method has higher recognition accuracy and adaptability to errors than traditional statistical methods and general deep neural networks,and it has good generalization for the recognition of new targets.(4)For the identification of radar radiation source models in complex scenarios,including low intercept probability technology and uneven observation information caused by radar periodic rotation,and the same or similar modes of multifunctional radar and complex mode radar,the paper proposes based on Attention-based Multi-RNNs(ABMR)model recognition method of attention mechanism.The model first assigns recurrent neural networks to the multi-dimensional characteristics of the radiation source,and uses the attention mechanism to reasonably allocate the weight of the neural network.During the recognition process,it focuses on the real pulse information that is helpful for recognition,and ignores noise and unobserved relevant information.The ABMR model uses high-dimensional features in pulse description words(Pulse Repetitive Words,PDWs)as network inputs for training,and the converged network can identify the acquired pulse sequences.Because the attention mechanism can focus on the key information of the pulse sequence by adjusting the network weights,it can fully mine and analyze the small differences in the characteristics of the radiation source and the less real pulses in the highnoise environment,so as to carry out the radar radiation source recognition in complex scene.Experiments show that the recognition performance of the new method under low signal-to-noise ratio is far superior to the existing traditional methods and general neural network methods.